1 /* 2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "gc/shared/barrierSet.hpp" 26 #include "gc/shared/c2/barrierSetC2.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/block.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealGraphPrinter.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/opcodes.hpp" 38 #include "opto/phaseX.hpp" 39 #include "opto/regalloc.hpp" 40 #include "opto/rootnode.hpp" 41 #include "utilities/macros.hpp" 42 #include "utilities/powerOfTwo.hpp" 43 44 //============================================================================= 45 #define NODE_HASH_MINIMUM_SIZE 255 46 47 //------------------------------NodeHash--------------------------------------- 48 NodeHash::NodeHash(Arena *arena, uint est_max_size) : 49 _a(arena), 50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ), 51 _inserts(0), _insert_limit( insert_limit() ), 52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ) 53 #ifndef PRODUCT 54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0), 55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0), 56 _total_inserts(0), _total_insert_probes(0) 57 #endif 58 { 59 // _sentinel must be in the current node space 60 _sentinel = new ProjNode(nullptr, TypeFunc::Control); 61 memset(_table,0,sizeof(Node*)*_max); 62 } 63 64 //------------------------------hash_find-------------------------------------- 65 // Find in hash table 66 Node *NodeHash::hash_find( const Node *n ) { 67 // ((Node*)n)->set_hash( n->hash() ); 68 uint hash = n->hash(); 69 if (hash == Node::NO_HASH) { 70 NOT_PRODUCT( _lookup_misses++ ); 71 return nullptr; 72 } 73 uint key = hash & (_max-1); 74 uint stride = key | 0x01; 75 NOT_PRODUCT( _look_probes++ ); 76 Node *k = _table[key]; // Get hashed value 77 if( !k ) { // ?Miss? 78 NOT_PRODUCT( _lookup_misses++ ); 79 return nullptr; // Miss! 80 } 81 82 int op = n->Opcode(); 83 uint req = n->req(); 84 while( 1 ) { // While probing hash table 85 if( k->req() == req && // Same count of inputs 86 k->Opcode() == op ) { // Same Opcode 87 for( uint i=0; i<req; i++ ) 88 if( n->in(i)!=k->in(i)) // Different inputs? 89 goto collision; // "goto" is a speed hack... 90 if( n->cmp(*k) ) { // Check for any special bits 91 NOT_PRODUCT( _lookup_hits++ ); 92 return k; // Hit! 93 } 94 } 95 collision: 96 NOT_PRODUCT( _look_probes++ ); 97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime 98 k = _table[key]; // Get hashed value 99 if( !k ) { // ?Miss? 100 NOT_PRODUCT( _lookup_misses++ ); 101 return nullptr; // Miss! 102 } 103 } 104 ShouldNotReachHere(); 105 return nullptr; 106 } 107 108 //------------------------------hash_find_insert------------------------------- 109 // Find in hash table, insert if not already present 110 // Used to preserve unique entries in hash table 111 Node *NodeHash::hash_find_insert( Node *n ) { 112 // n->set_hash( ); 113 uint hash = n->hash(); 114 if (hash == Node::NO_HASH) { 115 NOT_PRODUCT( _lookup_misses++ ); 116 return nullptr; 117 } 118 uint key = hash & (_max-1); 119 uint stride = key | 0x01; // stride must be relatively prime to table siz 120 uint first_sentinel = 0; // replace a sentinel if seen. 121 NOT_PRODUCT( _look_probes++ ); 122 Node *k = _table[key]; // Get hashed value 123 if( !k ) { // ?Miss? 124 NOT_PRODUCT( _lookup_misses++ ); 125 _table[key] = n; // Insert into table! 126 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 127 check_grow(); // Grow table if insert hit limit 128 return nullptr; // Miss! 129 } 130 else if( k == _sentinel ) { 131 first_sentinel = key; // Can insert here 132 } 133 134 int op = n->Opcode(); 135 uint req = n->req(); 136 while( 1 ) { // While probing hash table 137 if( k->req() == req && // Same count of inputs 138 k->Opcode() == op ) { // Same Opcode 139 for( uint i=0; i<req; i++ ) 140 if( n->in(i)!=k->in(i)) // Different inputs? 141 goto collision; // "goto" is a speed hack... 142 if( n->cmp(*k) ) { // Check for any special bits 143 NOT_PRODUCT( _lookup_hits++ ); 144 return k; // Hit! 145 } 146 } 147 collision: 148 NOT_PRODUCT( _look_probes++ ); 149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 150 k = _table[key]; // Get hashed value 151 if( !k ) { // ?Miss? 152 NOT_PRODUCT( _lookup_misses++ ); 153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel? 154 _table[key] = n; // Insert into table! 155 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 156 check_grow(); // Grow table if insert hit limit 157 return nullptr; // Miss! 158 } 159 else if( first_sentinel == 0 && k == _sentinel ) { 160 first_sentinel = key; // Can insert here 161 } 162 163 } 164 ShouldNotReachHere(); 165 return nullptr; 166 } 167 168 //------------------------------hash_insert------------------------------------ 169 // Insert into hash table 170 void NodeHash::hash_insert( Node *n ) { 171 // // "conflict" comments -- print nodes that conflict 172 // bool conflict = false; 173 // n->set_hash(); 174 uint hash = n->hash(); 175 if (hash == Node::NO_HASH) { 176 return; 177 } 178 check_grow(); 179 uint key = hash & (_max-1); 180 uint stride = key | 0x01; 181 182 while( 1 ) { // While probing hash table 183 NOT_PRODUCT( _insert_probes++ ); 184 Node *k = _table[key]; // Get hashed value 185 if( !k || (k == _sentinel) ) break; // Found a slot 186 assert( k != n, "already inserted" ); 187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; } 188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 189 } 190 _table[key] = n; // Insert into table! 191 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 192 // if( conflict ) { n->dump(); } 193 } 194 195 //------------------------------hash_delete------------------------------------ 196 // Replace in hash table with sentinel 197 bool NodeHash::hash_delete( const Node *n ) { 198 Node *k; 199 uint hash = n->hash(); 200 if (hash == Node::NO_HASH) { 201 NOT_PRODUCT( _delete_misses++ ); 202 return false; 203 } 204 uint key = hash & (_max-1); 205 uint stride = key | 0x01; 206 debug_only( uint counter = 0; ); 207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) { 208 debug_only( counter++ ); 209 NOT_PRODUCT( _delete_probes++ ); 210 k = _table[key]; // Get hashed value 211 if( !k ) { // Miss? 212 NOT_PRODUCT( _delete_misses++ ); 213 return false; // Miss! Not in chain 214 } 215 else if( n == k ) { 216 NOT_PRODUCT( _delete_hits++ ); 217 _table[key] = _sentinel; // Hit! Label as deleted entry 218 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table. 219 return true; 220 } 221 else { 222 // collision: move through table with prime offset 223 key = (key + stride/*7*/) & (_max-1); 224 assert( counter <= _insert_limit, "Cycle in hash-table"); 225 } 226 } 227 ShouldNotReachHere(); 228 return false; 229 } 230 231 //------------------------------round_up--------------------------------------- 232 // Round up to nearest power of 2 233 uint NodeHash::round_up(uint x) { 234 x += (x >> 2); // Add 25% slop 235 return MAX2(16U, round_up_power_of_2(x)); 236 } 237 238 //------------------------------grow------------------------------------------- 239 // Grow _table to next power of 2 and insert old entries 240 void NodeHash::grow() { 241 // Record old state 242 uint old_max = _max; 243 Node **old_table = _table; 244 // Construct new table with twice the space 245 #ifndef PRODUCT 246 _grows++; 247 _total_inserts += _inserts; 248 _total_insert_probes += _insert_probes; 249 _insert_probes = 0; 250 #endif 251 _inserts = 0; 252 _max = _max << 1; 253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) ); 254 memset(_table,0,sizeof(Node*)*_max); 255 _insert_limit = insert_limit(); 256 // Insert old entries into the new table 257 for( uint i = 0; i < old_max; i++ ) { 258 Node *m = *old_table++; 259 if( !m || m == _sentinel ) continue; 260 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table. 261 hash_insert(m); 262 } 263 } 264 265 //------------------------------clear------------------------------------------ 266 // Clear all entries in _table to null but keep storage 267 void NodeHash::clear() { 268 #ifdef ASSERT 269 // Unlock all nodes upon removal from table. 270 for (uint i = 0; i < _max; i++) { 271 Node* n = _table[i]; 272 if (!n || n == _sentinel) continue; 273 n->exit_hash_lock(); 274 } 275 #endif 276 277 memset( _table, 0, _max * sizeof(Node*) ); 278 } 279 280 //-----------------------remove_useless_nodes---------------------------------- 281 // Remove useless nodes from value table, 282 // implementation does not depend on hash function 283 void NodeHash::remove_useless_nodes(VectorSet &useful) { 284 285 // Dead nodes in the hash table inherited from GVN should not replace 286 // existing nodes, remove dead nodes. 287 uint max = size(); 288 Node *sentinel_node = sentinel(); 289 for( uint i = 0; i < max; ++i ) { 290 Node *n = at(i); 291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) { 292 debug_only(n->exit_hash_lock()); // Unlock the node when removed 293 _table[i] = sentinel_node; // Replace with placeholder 294 } 295 } 296 } 297 298 299 void NodeHash::check_no_speculative_types() { 300 #ifdef ASSERT 301 uint max = size(); 302 Unique_Node_List live_nodes; 303 Compile::current()->identify_useful_nodes(live_nodes); 304 Node *sentinel_node = sentinel(); 305 for (uint i = 0; i < max; ++i) { 306 Node *n = at(i); 307 if (n != nullptr && 308 n != sentinel_node && 309 n->is_Type() && 310 live_nodes.member(n)) { 311 TypeNode* tn = n->as_Type(); 312 const Type* t = tn->type(); 313 const Type* t_no_spec = t->remove_speculative(); 314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup"); 315 } 316 } 317 #endif 318 } 319 320 #ifndef PRODUCT 321 //------------------------------dump------------------------------------------- 322 // Dump statistics for the hash table 323 void NodeHash::dump() { 324 _total_inserts += _inserts; 325 _total_insert_probes += _insert_probes; 326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) { 327 if (WizardMode) { 328 for (uint i=0; i<_max; i++) { 329 if (_table[i]) 330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx); 331 } 332 } 333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max); 334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0); 335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses)); 336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts); 337 // sentinels increase lookup cost, but not insert cost 338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function"); 339 assert( _inserts+(_inserts>>3) < _max, "table too full" ); 340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" ); 341 } 342 } 343 344 Node *NodeHash::find_index(uint idx) { // For debugging 345 // Find an entry by its index value 346 for( uint i = 0; i < _max; i++ ) { 347 Node *m = _table[i]; 348 if( !m || m == _sentinel ) continue; 349 if( m->_idx == (uint)idx ) return m; 350 } 351 return nullptr; 352 } 353 #endif 354 355 #ifdef ASSERT 356 NodeHash::~NodeHash() { 357 // Unlock all nodes upon destruction of table. 358 if (_table != (Node**)badAddress) clear(); 359 } 360 #endif 361 362 363 //============================================================================= 364 //------------------------------PhaseRemoveUseless----------------------------- 365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root. 366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) { 367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 368 // Implementation requires an edge from root to each SafePointNode 369 // at a backward branch. Inserted in add_safepoint(). 370 371 // Identify nodes that are reachable from below, useful. 372 C->identify_useful_nodes(_useful); 373 // Update dead node list 374 C->update_dead_node_list(_useful); 375 376 // Remove all useless nodes from PhaseValues' recorded types 377 // Must be done before disconnecting nodes to preserve hash-table-invariant 378 gvn->remove_useless_nodes(_useful.member_set()); 379 380 // Remove all useless nodes from future worklist 381 worklist.remove_useless_nodes(_useful.member_set()); 382 383 // Disconnect 'useless' nodes that are adjacent to useful nodes 384 C->disconnect_useless_nodes(_useful, worklist); 385 } 386 387 //============================================================================= 388 //------------------------------PhaseRenumberLive------------------------------ 389 // First, remove useless nodes (equivalent to identifying live nodes). 390 // Then, renumber live nodes. 391 // 392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure. 393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the 394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique 395 // value in the range [0, x). 396 // 397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is 398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes). 399 // 400 // The PhaseRenumberLive phase updates two data structures with the new node IDs. 401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to 402 // be processed by IGVN after removal of the useless nodes. 403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to 404 // the node's type. The mapping is updated to use the new node IDs as well. We 405 // create a new map, and swap it with the old one. 406 // 407 // Other data structures used by the compiler are not updated. The hash table for value 408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because 409 // computing the hash values is not based on node IDs. 410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn, 411 Unique_Node_List& worklist, 412 PhaseNumber phase_num) : 413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live), 414 _new_type_array(C->comp_arena()), 415 _old2new_map(C->unique(), C->unique(), -1), 416 _is_pass_finished(false), 417 _live_node_count(C->live_nodes()) 418 { 419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place"); 420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes"); 421 assert(_delayed.size() == 0, "should be empty"); 422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile"); 423 assert(&gvn->types() == C->types(), "reference still same as that from Compile"); 424 425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array(); 426 if (old_node_note_array != nullptr) { 427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size 428 new_size = MAX2(8, new_size); 429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr)); 430 C->grow_node_notes(C->node_note_array(), new_size); 431 } 432 433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist"); 434 435 // Iterate over the set of live nodes. 436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) { 437 Node* n = _useful.at(current_idx); 438 439 const Type* type = gvn->type_or_null(n); 440 _new_type_array.map(current_idx, type); 441 442 assert(_old2new_map.at(n->_idx) == -1, "already seen"); 443 _old2new_map.at_put(n->_idx, current_idx); 444 445 if (old_node_note_array != nullptr) { 446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx); 447 C->set_node_notes_at(current_idx, nn); 448 } 449 450 n->set_idx(current_idx); // Update node ID. 451 452 if (update_embedded_ids(n) < 0) { 453 _delayed.push(n); // has embedded IDs; handle later 454 } 455 } 456 457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed. 458 worklist.recompute_idx_set(); 459 460 assert(_live_node_count == _useful.size(), "all live nodes must be processed"); 461 462 _is_pass_finished = true; // pass finished; safe to process delayed updates 463 464 while (_delayed.size() > 0) { 465 Node* n = _delayed.pop(); 466 int no_of_updates = update_embedded_ids(n); 467 assert(no_of_updates > 0, "should be updated"); 468 } 469 470 // Replace the compiler's type information with the updated type information. 471 gvn->types().swap(_new_type_array); 472 473 // Update the unique node count of the compilation to the number of currently live nodes. 474 C->set_unique(_live_node_count); 475 476 // Set the dead node count to 0 and reset dead node list. 477 C->reset_dead_node_list(); 478 } 479 480 int PhaseRenumberLive::new_index(int old_idx) { 481 assert(_is_pass_finished, "not finished"); 482 if (_old2new_map.at(old_idx) == -1) { // absent 483 // Allocate a placeholder to preserve uniqueness 484 _old2new_map.at_put(old_idx, _live_node_count); 485 _live_node_count++; 486 } 487 return _old2new_map.at(old_idx); 488 } 489 490 int PhaseRenumberLive::update_embedded_ids(Node* n) { 491 int no_of_updates = 0; 492 if (n->is_Phi()) { 493 PhiNode* phi = n->as_Phi(); 494 if (phi->_inst_id != -1) { 495 if (!_is_pass_finished) { 496 return -1; // delay 497 } 498 int new_idx = new_index(phi->_inst_id); 499 assert(new_idx != -1, ""); 500 phi->_inst_id = new_idx; 501 no_of_updates++; 502 } 503 if (phi->_inst_mem_id != -1) { 504 if (!_is_pass_finished) { 505 return -1; // delay 506 } 507 int new_idx = new_index(phi->_inst_mem_id); 508 assert(new_idx != -1, ""); 509 phi->_inst_mem_id = new_idx; 510 no_of_updates++; 511 } 512 } 513 514 const Type* type = _new_type_array.fast_lookup(n->_idx); 515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) { 516 if (!_is_pass_finished) { 517 return -1; // delay 518 } 519 int old_idx = type->is_oopptr()->instance_id(); 520 int new_idx = new_index(old_idx); 521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx); 522 _new_type_array.map(n->_idx, new_type); 523 no_of_updates++; 524 } 525 526 return no_of_updates; 527 } 528 529 void PhaseValues::init_con_caches() { 530 memset(_icons,0,sizeof(_icons)); 531 memset(_lcons,0,sizeof(_lcons)); 532 memset(_zcons,0,sizeof(_zcons)); 533 } 534 535 //--------------------------------find_int_type-------------------------------- 536 const TypeInt* PhaseValues::find_int_type(Node* n) { 537 if (n == nullptr) return nullptr; 538 // Call type_or_null(n) to determine node's type since we might be in 539 // parse phase and call n->Value() may return wrong type. 540 // (For example, a phi node at the beginning of loop parsing is not ready.) 541 const Type* t = type_or_null(n); 542 if (t == nullptr) return nullptr; 543 return t->isa_int(); 544 } 545 546 547 //-------------------------------find_long_type-------------------------------- 548 const TypeLong* PhaseValues::find_long_type(Node* n) { 549 if (n == nullptr) return nullptr; 550 // (See comment above on type_or_null.) 551 const Type* t = type_or_null(n); 552 if (t == nullptr) return nullptr; 553 return t->isa_long(); 554 } 555 556 //------------------------------~PhaseValues----------------------------------- 557 #ifndef PRODUCT 558 PhaseValues::~PhaseValues() { 559 // Statistics for NodeHash 560 _table.dump(); 561 // Statistics for value progress and efficiency 562 if( PrintCompilation && Verbose && WizardMode ) { 563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 565 if( made_transforms() != 0 ) { 566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 567 } else { 568 tty->cr(); 569 } 570 } 571 } 572 #endif 573 574 //------------------------------makecon---------------------------------------- 575 ConNode* PhaseValues::makecon(const Type* t) { 576 assert(t->singleton(), "must be a constant"); 577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 578 switch (t->base()) { // fast paths 579 case Type::Half: 580 case Type::Top: return (ConNode*) C->top(); 581 case Type::Int: return intcon( t->is_int()->get_con() ); 582 case Type::Long: return longcon( t->is_long()->get_con() ); 583 default: break; 584 } 585 if (t->is_zero_type()) 586 return zerocon(t->basic_type()); 587 return uncached_makecon(t); 588 } 589 590 //--------------------------uncached_makecon----------------------------------- 591 // Make an idealized constant - one of ConINode, ConPNode, etc. 592 ConNode* PhaseValues::uncached_makecon(const Type *t) { 593 assert(t->singleton(), "must be a constant"); 594 ConNode* x = ConNode::make(t); 595 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 596 if (k == nullptr) { 597 set_type(x, t); // Missed, provide type mapping 598 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 599 if (nna != nullptr) { 600 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 601 loc->clear(); // do not put debug info on constants 602 } 603 } else { 604 x->destruct(this); // Hit, destroy duplicate constant 605 x = k; // use existing constant 606 } 607 return x; 608 } 609 610 //------------------------------intcon----------------------------------------- 611 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 612 ConINode* PhaseValues::intcon(jint i) { 613 // Small integer? Check cache! Check that cached node is not dead 614 if (i >= _icon_min && i <= _icon_max) { 615 ConINode* icon = _icons[i-_icon_min]; 616 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr) 617 return icon; 618 } 619 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 620 assert(icon->is_Con(), ""); 621 if (i >= _icon_min && i <= _icon_max) 622 _icons[i-_icon_min] = icon; // Cache small integers 623 return icon; 624 } 625 626 //------------------------------longcon---------------------------------------- 627 // Fast long constant. 628 ConLNode* PhaseValues::longcon(jlong l) { 629 // Small integer? Check cache! Check that cached node is not dead 630 if (l >= _lcon_min && l <= _lcon_max) { 631 ConLNode* lcon = _lcons[l-_lcon_min]; 632 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr) 633 return lcon; 634 } 635 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 636 assert(lcon->is_Con(), ""); 637 if (l >= _lcon_min && l <= _lcon_max) 638 _lcons[l-_lcon_min] = lcon; // Cache small integers 639 return lcon; 640 } 641 ConNode* PhaseValues::integercon(jlong l, BasicType bt) { 642 if (bt == T_INT) { 643 return intcon(checked_cast<jint>(l)); 644 } 645 assert(bt == T_LONG, "not an integer"); 646 return longcon(l); 647 } 648 649 650 //------------------------------zerocon----------------------------------------- 651 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 652 ConNode* PhaseValues::zerocon(BasicType bt) { 653 assert((uint)bt <= _zcon_max, "domain check"); 654 ConNode* zcon = _zcons[bt]; 655 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr) 656 return zcon; 657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 658 _zcons[bt] = zcon; 659 return zcon; 660 } 661 662 663 664 //============================================================================= 665 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) { 666 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape); 667 if (i == nullptr) { 668 i = k->Ideal(this, can_reshape); 669 } 670 return i; 671 } 672 673 //------------------------------transform-------------------------------------- 674 // Return a node which computes the same function as this node, but 675 // in a faster or cheaper fashion. 676 Node* PhaseGVN::transform(Node* n) { 677 NOT_PRODUCT( set_transforms(); ) 678 679 // Apply the Ideal call in a loop until it no longer applies 680 Node* k = n; 681 Node* i = apply_ideal(k, /*can_reshape=*/false); 682 NOT_PRODUCT(uint loop_count = 1;) 683 while (i != nullptr) { 684 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 685 k = i; 686 #ifdef ASSERT 687 if (loop_count >= K + C->live_nodes()) { 688 dump_infinite_loop_info(i, "PhaseGVN::transform"); 689 } 690 #endif 691 i = apply_ideal(k, /*can_reshape=*/false); 692 NOT_PRODUCT(loop_count++;) 693 } 694 NOT_PRODUCT(if (loop_count != 0) { set_progress(); }) 695 696 // If brand new node, make space in type array. 697 ensure_type_or_null(k); 698 699 // Since I just called 'Value' to compute the set of run-time values 700 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 701 // cache Value. Later requests for the local phase->type of this Node can 702 // use the cached Value instead of suffering with 'bottom_type'. 703 const Type* t = k->Value(this); // Get runtime Value set 704 assert(t != nullptr, "value sanity"); 705 if (type_or_null(k) != t) { 706 #ifndef PRODUCT 707 // Do not count initial visit to node as a transformation 708 if (type_or_null(k) == nullptr) { 709 inc_new_values(); 710 set_progress(); 711 } 712 #endif 713 set_type(k, t); 714 // If k is a TypeNode, capture any more-precise type permanently into Node 715 k->raise_bottom_type(t); 716 } 717 718 if (t->singleton() && !k->is_Con()) { 719 NOT_PRODUCT(set_progress();) 720 return makecon(t); // Turn into a constant 721 } 722 723 // Now check for Identities 724 i = k->Identity(this); // Look for a nearby replacement 725 if (i != k) { // Found? Return replacement! 726 NOT_PRODUCT(set_progress();) 727 return i; 728 } 729 730 // Global Value Numbering 731 i = hash_find_insert(k); // Insert if new 732 if (i && (i != k)) { 733 // Return the pre-existing node 734 NOT_PRODUCT(set_progress();) 735 return i; 736 } 737 738 // Return Idealized original 739 return k; 740 } 741 742 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 743 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) { 744 return false; 745 } 746 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) { 747 return false; 748 } 749 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 750 int i = 0; 751 while (d != n) { 752 n = IfNode::up_one_dom(n, linear_only); 753 i++; 754 if (n == nullptr || i >= 100) { 755 return false; 756 } 757 } 758 return true; 759 } 760 761 #ifdef ASSERT 762 //------------------------------dead_loop_check-------------------------------- 763 // Check for a simple dead loop when a data node references itself directly 764 // or through an other data node excluding cons and phis. 765 void PhaseGVN::dead_loop_check( Node *n ) { 766 // Phi may reference itself in a loop 767 if (n != nullptr && !n->is_dead_loop_safe() && !n->is_CFG()) { 768 // Do 2 levels check and only data inputs. 769 bool no_dead_loop = true; 770 uint cnt = n->req(); 771 for (uint i = 1; i < cnt && no_dead_loop; i++) { 772 Node *in = n->in(i); 773 if (in == n) { 774 no_dead_loop = false; 775 } else if (in != nullptr && !in->is_dead_loop_safe()) { 776 uint icnt = in->req(); 777 for (uint j = 1; j < icnt && no_dead_loop; j++) { 778 if (in->in(j) == n || in->in(j) == in) 779 no_dead_loop = false; 780 } 781 } 782 } 783 if (!no_dead_loop) n->dump_bfs(100,nullptr,"#"); 784 assert(no_dead_loop, "dead loop detected"); 785 } 786 } 787 788 789 /** 790 * Dumps information that can help to debug the problem. A debug 791 * build fails with an assert. 792 */ 793 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) { 794 n->dump(4); 795 assert(false, "infinite loop in %s", where); 796 } 797 #endif 798 799 //============================================================================= 800 //------------------------------PhaseIterGVN----------------------------------- 801 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 802 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform), 803 _worklist(*C->igvn_worklist()) 804 { 805 _iterGVN = true; 806 assert(&_worklist == &igvn->_worklist, "sanity"); 807 } 808 809 //------------------------------PhaseIterGVN----------------------------------- 810 // Initialize with previous PhaseGVN info from Parser 811 PhaseIterGVN::PhaseIterGVN(PhaseGVN* gvn) : _delay_transform(false), 812 _worklist(*C->igvn_worklist()) 813 { 814 _iterGVN = true; 815 uint max; 816 817 // Dead nodes in the hash table inherited from GVN were not treated as 818 // roots during def-use info creation; hence they represent an invisible 819 // use. Clear them out. 820 max = _table.size(); 821 for( uint i = 0; i < max; ++i ) { 822 Node *n = _table.at(i); 823 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) { 824 if( n->is_top() ) continue; 825 // If remove_useless_nodes() has run, we expect no such nodes left. 826 assert(false, "remove_useless_nodes missed this node"); 827 hash_delete(n); 828 } 829 } 830 831 // Any Phis or Regions on the worklist probably had uses that could not 832 // make more progress because the uses were made while the Phis and Regions 833 // were in half-built states. Put all uses of Phis and Regions on worklist. 834 max = _worklist.size(); 835 for( uint j = 0; j < max; j++ ) { 836 Node *n = _worklist.at(j); 837 uint uop = n->Opcode(); 838 if( uop == Op_Phi || uop == Op_Region || 839 n->is_Type() || 840 n->is_Mem() ) 841 add_users_to_worklist(n); 842 } 843 } 844 845 void PhaseIterGVN::shuffle_worklist() { 846 if (_worklist.size() < 2) return; 847 for (uint i = _worklist.size() - 1; i >= 1; i--) { 848 uint j = C->random() % (i + 1); 849 swap(_worklist.adr()[i], _worklist.adr()[j]); 850 } 851 } 852 853 #ifndef PRODUCT 854 void PhaseIterGVN::verify_step(Node* n) { 855 if (is_verify_def_use()) { 856 ResourceMark rm; 857 VectorSet visited; 858 Node_List worklist; 859 860 _verify_window[_verify_counter % _verify_window_size] = n; 861 ++_verify_counter; 862 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 863 ++_verify_full_passes; 864 worklist.push(C->root()); 865 Node::verify(-1, visited, worklist); 866 return; 867 } 868 for (int i = 0; i < _verify_window_size; i++) { 869 Node* n = _verify_window[i]; 870 if (n == nullptr) { 871 continue; 872 } 873 if (n->in(0) == NodeSentinel) { // xform_idom 874 _verify_window[i] = n->in(1); 875 --i; 876 continue; 877 } 878 // Typical fanout is 1-2, so this call visits about 6 nodes. 879 if (!visited.test_set(n->_idx)) { 880 worklist.push(n); 881 } 882 } 883 Node::verify(4, visited, worklist); 884 } 885 } 886 887 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 888 const Type* newtype = type_or_null(n); 889 if (nn != n || oldtype != newtype) { 890 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n); 891 } 892 if (TraceIterativeGVN) { 893 uint wlsize = _worklist.size(); 894 if (nn != n) { 895 // print old node 896 tty->print("< "); 897 if (oldtype != newtype && oldtype != nullptr) { 898 oldtype->dump(); 899 } 900 do { tty->print("\t"); } while (tty->position() < 16); 901 tty->print("<"); 902 n->dump(); 903 } 904 if (oldtype != newtype || nn != n) { 905 // print new node and/or new type 906 if (oldtype == nullptr) { 907 tty->print("* "); 908 } else if (nn != n) { 909 tty->print("> "); 910 } else { 911 tty->print("= "); 912 } 913 if (newtype == nullptr) { 914 tty->print("null"); 915 } else { 916 newtype->dump(); 917 } 918 do { tty->print("\t"); } while (tty->position() < 16); 919 nn->dump(); 920 } 921 if (Verbose && wlsize < _worklist.size()) { 922 tty->print(" Push {"); 923 while (wlsize != _worklist.size()) { 924 Node* pushed = _worklist.at(wlsize++); 925 tty->print(" %d", pushed->_idx); 926 } 927 tty->print_cr(" }"); 928 } 929 if (nn != n) { 930 // ignore n, it might be subsumed 931 verify_step((Node*) nullptr); 932 } 933 } 934 } 935 936 void PhaseIterGVN::init_verifyPhaseIterGVN() { 937 _verify_counter = 0; 938 _verify_full_passes = 0; 939 for (int i = 0; i < _verify_window_size; i++) { 940 _verify_window[i] = nullptr; 941 } 942 #ifdef ASSERT 943 // Verify that all modified nodes are on _worklist 944 Unique_Node_List* modified_list = C->modified_nodes(); 945 while (modified_list != nullptr && modified_list->size()) { 946 Node* n = modified_list->pop(); 947 if (!n->is_Con() && !_worklist.member(n)) { 948 n->dump(); 949 fatal("modified node is not on IGVN._worklist"); 950 } 951 } 952 #endif 953 } 954 955 void PhaseIterGVN::verify_PhaseIterGVN() { 956 #ifdef ASSERT 957 // Verify nodes with changed inputs. 958 Unique_Node_List* modified_list = C->modified_nodes(); 959 while (modified_list != nullptr && modified_list->size()) { 960 Node* n = modified_list->pop(); 961 if (!n->is_Con()) { // skip Con nodes 962 n->dump(); 963 fatal("modified node was not processed by IGVN.transform_old()"); 964 } 965 } 966 #endif 967 968 C->verify_graph_edges(); 969 if (is_verify_def_use() && PrintOpto) { 970 if (_verify_counter == _verify_full_passes) { 971 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 972 (int) _verify_full_passes); 973 } else { 974 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 975 (int) _verify_counter, (int) _verify_full_passes); 976 } 977 } 978 979 #ifdef ASSERT 980 if (modified_list != nullptr) { 981 while (modified_list->size() > 0) { 982 Node* n = modified_list->pop(); 983 n->dump(); 984 assert(false, "VerifyIterativeGVN: new modified node was added"); 985 } 986 } 987 988 verify_optimize(); 989 #endif 990 } 991 #endif /* PRODUCT */ 992 993 #ifdef ASSERT 994 /** 995 * Dumps information that can help to debug the problem. A debug 996 * build fails with an assert. 997 */ 998 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) { 999 n->dump(4); 1000 _worklist.dump(); 1001 assert(false, "infinite loop in %s", where); 1002 } 1003 1004 /** 1005 * Prints out information about IGVN if the 'verbose' option is used. 1006 */ 1007 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1008 if (TraceIterativeGVN && Verbose) { 1009 tty->print(" Pop "); 1010 n->dump(); 1011 if ((num_processed % 100) == 0) { 1012 _worklist.print_set(); 1013 } 1014 } 1015 } 1016 #endif /* ASSERT */ 1017 1018 void PhaseIterGVN::optimize() { 1019 DEBUG_ONLY(uint num_processed = 0;) 1020 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1021 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);) 1022 C->print_method(PHASE_BEFORE_ITER_GVN, 3); 1023 if (StressIGVN) { 1024 shuffle_worklist(); 1025 } 1026 1027 // The node count check in the loop below (check_node_count) assumes that we 1028 // increase the live node count with at most 1029 // max_live_nodes_increase_per_iteration in between checks. If this 1030 // assumption does not hold, there is a risk that we exceed the max node 1031 // limit in between checks and trigger an assert during node creation. 1032 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 2; 1033 1034 uint loop_count = 0; 1035 // Pull from worklist and transform the node. If the node has changed, 1036 // update edge info and put uses on worklist. 1037 while (_worklist.size() > 0) { 1038 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) { 1039 C->print_method(PHASE_AFTER_ITER_GVN, 3); 1040 return; 1041 } 1042 Node* n = _worklist.pop(); 1043 if (loop_count >= K * C->live_nodes()) { 1044 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");) 1045 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1046 C->print_method(PHASE_AFTER_ITER_GVN, 3); 1047 return; 1048 } 1049 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1050 if (n->outcnt() != 0) { 1051 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1052 // Do the transformation 1053 DEBUG_ONLY(int live_nodes_before = C->live_nodes();) 1054 Node* nn = transform_old(n); 1055 DEBUG_ONLY(int live_nodes_after = C->live_nodes();) 1056 // Ensure we did not increase the live node count with more than 1057 // max_live_nodes_increase_per_iteration during the call to transform_old 1058 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;) 1059 assert(increase < max_live_nodes_increase_per_iteration, 1060 "excessive live node increase in single iteration of IGVN: %d " 1061 "(should be at most %d)", 1062 increase, max_live_nodes_increase_per_iteration); 1063 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1064 } else if (!n->is_top()) { 1065 remove_dead_node(n); 1066 } 1067 loop_count++; 1068 } 1069 NOT_PRODUCT(verify_PhaseIterGVN();) 1070 C->print_method(PHASE_AFTER_ITER_GVN, 3); 1071 } 1072 1073 #ifdef ASSERT 1074 void PhaseIterGVN::verify_optimize() { 1075 if (is_verify_Value()) { 1076 ResourceMark rm; 1077 Unique_Node_List worklist; 1078 bool failure = false; 1079 // BFS all nodes, starting at root 1080 worklist.push(C->root()); 1081 for (uint j = 0; j < worklist.size(); ++j) { 1082 Node* n = worklist.at(j); 1083 failure |= verify_node_value(n); 1084 // traverse all inputs and outputs 1085 for (uint i = 0; i < n->req(); i++) { 1086 if (n->in(i) != nullptr) { 1087 worklist.push(n->in(i)); 1088 } 1089 } 1090 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1091 worklist.push(n->fast_out(i)); 1092 } 1093 } 1094 // If we get this assert, check why the reported nodes were not processed again in IGVN. 1095 // We should either make sure that these nodes are properly added back to the IGVN worklist 1096 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception 1097 // in the verification code above if that is not possible for some reason (like Load nodes). 1098 assert(!failure, "Missed optimization opportunity in PhaseIterGVN"); 1099 } 1100 } 1101 1102 // Check that type(n) == n->Value(), return true if we have a failure. 1103 // We have a list of exceptions, see detailed comments in code. 1104 // (1) Integer "widen" changes, but the range is the same. 1105 // (2) LoadNode performs deep traversals. Load is not notified for changes far away. 1106 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away. 1107 bool PhaseIterGVN::verify_node_value(Node* n) { 1108 // If we assert inside type(n), because the type is still a null, then maybe 1109 // the node never went through gvn.transform, which would be a bug. 1110 const Type* told = type(n); 1111 const Type* tnew = n->Value(this); 1112 if (told == tnew) { 1113 return false; 1114 } 1115 // Exception (1) 1116 // Integer "widen" changes, but range is the same. 1117 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long 1118 const TypeInteger* t0 = told->is_integer(tnew->basic_type()); 1119 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type()); 1120 if (t0->lo_as_long() == t1->lo_as_long() && 1121 t0->hi_as_long() == t1->hi_as_long()) { 1122 return false; // ignore integer widen 1123 } 1124 } 1125 // Exception (2) 1126 // LoadNode performs deep traversals. Load is not notified for changes far away. 1127 if (n->is_Load() && !told->singleton()) { 1128 // MemNode::can_see_stored_value looks up through many memory nodes, 1129 // which means we would need to notify modifications from far up in 1130 // the inputs all the way down to the LoadNode. We don't do that. 1131 return false; 1132 } 1133 // Exception (3) 1134 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away. 1135 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) { 1136 // SubNode::Value 1137 // CmpPNode::sub 1138 // MemNode::detect_ptr_independence 1139 // MemNode::all_controls_dominate 1140 // We find all controls of a pointer load, and see if they dominate the control of 1141 // an allocation. If they all dominate, we know the allocation is after (independent) 1142 // of the pointer load, and we can say the pointers are different. For this we call 1143 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the 1144 // control sub of the allocation. The problems is that sometimes dominates answers 1145 // false conservatively, and later it can determine that it is indeed true. Loops with 1146 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and 1147 // so the traversal becomes more powerful. This is difficult to remidy, we would have 1148 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP 1149 // after loop-opts, so that should take care of many of these cases. 1150 return false; 1151 } 1152 tty->cr(); 1153 tty->print_cr("Missed Value optimization:"); 1154 n->dump_bfs(3, nullptr, ""); 1155 tty->print_cr("Current type:"); 1156 told->dump_on(tty); 1157 tty->cr(); 1158 tty->print_cr("Optimized type:"); 1159 tnew->dump_on(tty); 1160 tty->cr(); 1161 return true; 1162 } 1163 #endif 1164 1165 /** 1166 * Register a new node with the optimizer. Update the types array, the def-use 1167 * info. Put on worklist. 1168 */ 1169 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1170 set_type_bottom(n); 1171 _worklist.push(n); 1172 if (orig != nullptr) C->copy_node_notes_to(n, orig); 1173 return n; 1174 } 1175 1176 //------------------------------transform-------------------------------------- 1177 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1178 Node *PhaseIterGVN::transform( Node *n ) { 1179 // If brand new node, make space in type array, and give it a type. 1180 ensure_type_or_null(n); 1181 if (type_or_null(n) == nullptr) { 1182 set_type_bottom(n); 1183 } 1184 1185 if (_delay_transform) { 1186 // Add the node to the worklist but don't optimize for now 1187 _worklist.push(n); 1188 return n; 1189 } 1190 1191 return transform_old(n); 1192 } 1193 1194 Node *PhaseIterGVN::transform_old(Node* n) { 1195 NOT_PRODUCT(set_transforms()); 1196 // Remove 'n' from hash table in case it gets modified 1197 _table.hash_delete(n); 1198 #ifdef ASSERT 1199 if (is_verify_def_use()) { 1200 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1201 } 1202 #endif 1203 1204 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool 1205 if (n->is_Cmp()) { 1206 add_users_to_worklist(n); 1207 } 1208 1209 // Apply the Ideal call in a loop until it no longer applies 1210 Node* k = n; 1211 DEBUG_ONLY(dead_loop_check(k);) 1212 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1213 C->remove_modified_node(k); 1214 Node* i = apply_ideal(k, /*can_reshape=*/true); 1215 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1216 #ifndef PRODUCT 1217 verify_step(k); 1218 #endif 1219 1220 DEBUG_ONLY(uint loop_count = 1;) 1221 while (i != nullptr) { 1222 #ifdef ASSERT 1223 if (loop_count >= K + C->live_nodes()) { 1224 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old"); 1225 } 1226 #endif 1227 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1228 // Made a change; put users of original Node on worklist 1229 add_users_to_worklist(k); 1230 // Replacing root of transform tree? 1231 if (k != i) { 1232 // Make users of old Node now use new. 1233 subsume_node(k, i); 1234 k = i; 1235 } 1236 DEBUG_ONLY(dead_loop_check(k);) 1237 // Try idealizing again 1238 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1239 C->remove_modified_node(k); 1240 i = apply_ideal(k, /*can_reshape=*/true); 1241 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1242 #ifndef PRODUCT 1243 verify_step(k); 1244 #endif 1245 DEBUG_ONLY(loop_count++;) 1246 } 1247 1248 // If brand new node, make space in type array. 1249 ensure_type_or_null(k); 1250 1251 // See what kind of values 'k' takes on at runtime 1252 const Type* t = k->Value(this); 1253 assert(t != nullptr, "value sanity"); 1254 1255 // Since I just called 'Value' to compute the set of run-time values 1256 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1257 // cache Value. Later requests for the local phase->type of this Node can 1258 // use the cached Value instead of suffering with 'bottom_type'. 1259 if (type_or_null(k) != t) { 1260 #ifndef PRODUCT 1261 inc_new_values(); 1262 set_progress(); 1263 #endif 1264 set_type(k, t); 1265 // If k is a TypeNode, capture any more-precise type permanently into Node 1266 k->raise_bottom_type(t); 1267 // Move users of node to worklist 1268 add_users_to_worklist(k); 1269 } 1270 // If 'k' computes a constant, replace it with a constant 1271 if (t->singleton() && !k->is_Con()) { 1272 NOT_PRODUCT(set_progress();) 1273 Node* con = makecon(t); // Make a constant 1274 add_users_to_worklist(k); 1275 subsume_node(k, con); // Everybody using k now uses con 1276 return con; 1277 } 1278 1279 // Now check for Identities 1280 i = k->Identity(this); // Look for a nearby replacement 1281 if (i != k) { // Found? Return replacement! 1282 NOT_PRODUCT(set_progress();) 1283 add_users_to_worklist(k); 1284 subsume_node(k, i); // Everybody using k now uses i 1285 return i; 1286 } 1287 1288 // Global Value Numbering 1289 i = hash_find_insert(k); // Check for pre-existing node 1290 if (i && (i != k)) { 1291 // Return the pre-existing node if it isn't dead 1292 NOT_PRODUCT(set_progress();) 1293 add_users_to_worklist(k); 1294 subsume_node(k, i); // Everybody using k now uses i 1295 return i; 1296 } 1297 1298 // Return Idealized original 1299 return k; 1300 } 1301 1302 //---------------------------------saturate------------------------------------ 1303 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1304 const Type* limit_type) const { 1305 return new_type->narrow(old_type); 1306 } 1307 1308 //------------------------------remove_globally_dead_node---------------------- 1309 // Kill a globally dead Node. All uses are also globally dead and are 1310 // aggressively trimmed. 1311 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1312 enum DeleteProgress { 1313 PROCESS_INPUTS, 1314 PROCESS_OUTPUTS 1315 }; 1316 ResourceMark rm; 1317 Node_Stack stack(32); 1318 stack.push(dead, PROCESS_INPUTS); 1319 1320 while (stack.is_nonempty()) { 1321 dead = stack.node(); 1322 if (dead->Opcode() == Op_SafePoint) { 1323 dead->as_SafePoint()->disconnect_from_root(this); 1324 } 1325 uint progress_state = stack.index(); 1326 assert(dead != C->root(), "killing root, eh?"); 1327 assert(!dead->is_top(), "add check for top when pushing"); 1328 NOT_PRODUCT( set_progress(); ) 1329 if (progress_state == PROCESS_INPUTS) { 1330 // After following inputs, continue to outputs 1331 stack.set_index(PROCESS_OUTPUTS); 1332 if (!dead->is_Con()) { // Don't kill cons but uses 1333 bool recurse = false; 1334 // Remove from hash table 1335 _table.hash_delete( dead ); 1336 // Smash all inputs to 'dead', isolating him completely 1337 for (uint i = 0; i < dead->req(); i++) { 1338 Node *in = dead->in(i); 1339 if (in != nullptr && in != C->top()) { // Points to something? 1340 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges 1341 assert((nrep > 0), "sanity"); 1342 if (in->outcnt() == 0) { // Made input go dead? 1343 stack.push(in, PROCESS_INPUTS); // Recursively remove 1344 recurse = true; 1345 } else if (in->outcnt() == 1 && 1346 in->has_special_unique_user()) { 1347 _worklist.push(in->unique_out()); 1348 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1349 if (in->Opcode() == Op_Region) { 1350 _worklist.push(in); 1351 } else if (in->is_Store()) { 1352 DUIterator_Fast imax, i = in->fast_outs(imax); 1353 _worklist.push(in->fast_out(i)); 1354 i++; 1355 if (in->outcnt() == 2) { 1356 _worklist.push(in->fast_out(i)); 1357 i++; 1358 } 1359 assert(!(i < imax), "sanity"); 1360 } 1361 } else if (dead->is_data_proj_of_pure_function(in)) { 1362 _worklist.push(in); 1363 } else { 1364 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in); 1365 } 1366 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1367 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) { 1368 // A Load that directly follows an InitializeNode is 1369 // going away. The Stores that follow are candidates 1370 // again to be captured by the InitializeNode. 1371 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1372 Node *n = in->fast_out(j); 1373 if (n->is_Store()) { 1374 _worklist.push(n); 1375 } 1376 } 1377 } 1378 } // if (in != nullptr && in != C->top()) 1379 } // for (uint i = 0; i < dead->req(); i++) 1380 if (recurse) { 1381 continue; 1382 } 1383 } // if (!dead->is_Con()) 1384 } // if (progress_state == PROCESS_INPUTS) 1385 1386 // Aggressively kill globally dead uses 1387 // (Rather than pushing all the outs at once, we push one at a time, 1388 // plus the parent to resume later, because of the indefinite number 1389 // of edge deletions per loop trip.) 1390 if (dead->outcnt() > 0) { 1391 // Recursively remove output edges 1392 stack.push(dead->raw_out(0), PROCESS_INPUTS); 1393 } else { 1394 // Finished disconnecting all input and output edges. 1395 stack.pop(); 1396 // Remove dead node from iterative worklist 1397 _worklist.remove(dead); 1398 C->remove_useless_node(dead); 1399 } 1400 } // while (stack.is_nonempty()) 1401 } 1402 1403 //------------------------------subsume_node----------------------------------- 1404 // Remove users from node 'old' and add them to node 'nn'. 1405 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1406 if (old->Opcode() == Op_SafePoint) { 1407 old->as_SafePoint()->disconnect_from_root(this); 1408 } 1409 assert( old != hash_find(old), "should already been removed" ); 1410 assert( old != C->top(), "cannot subsume top node"); 1411 // Copy debug or profile information to the new version: 1412 C->copy_node_notes_to(nn, old); 1413 // Move users of node 'old' to node 'nn' 1414 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1415 Node* use = old->last_out(i); // for each use... 1416 // use might need re-hashing (but it won't if it's a new node) 1417 rehash_node_delayed(use); 1418 // Update use-def info as well 1419 // We remove all occurrences of old within use->in, 1420 // so as to avoid rehashing any node more than once. 1421 // The hash table probe swamps any outer loop overhead. 1422 uint num_edges = 0; 1423 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1424 if (use->in(j) == old) { 1425 use->set_req(j, nn); 1426 ++num_edges; 1427 } 1428 } 1429 i -= num_edges; // we deleted 1 or more copies of this edge 1430 } 1431 1432 // Search for instance field data PhiNodes in the same region pointing to the old 1433 // memory PhiNode and update their instance memory ids to point to the new node. 1434 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) { 1435 Node* region = old->in(0); 1436 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1437 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1438 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) { 1439 phi->set_inst_mem_id((int)nn->_idx); 1440 } 1441 } 1442 } 1443 1444 // Smash all inputs to 'old', isolating him completely 1445 Node *temp = new Node(1); 1446 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1447 remove_dead_node( old ); 1448 temp->del_req(0); // Yank bogus edge 1449 if (nn != nullptr && nn->outcnt() == 0) { 1450 _worklist.push(nn); 1451 } 1452 #ifndef PRODUCT 1453 if (is_verify_def_use()) { 1454 for ( int i = 0; i < _verify_window_size; i++ ) { 1455 if ( _verify_window[i] == old ) 1456 _verify_window[i] = nn; 1457 } 1458 } 1459 #endif 1460 temp->destruct(this); // reuse the _idx of this little guy 1461 } 1462 1463 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) { 1464 assert(n != nullptr, "sanity"); 1465 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1466 Node* u = n->fast_out(i); 1467 if (u != n) { 1468 rehash_node_delayed(u); 1469 int nb = u->replace_edge(n, m); 1470 --i, imax -= nb; 1471 } 1472 } 1473 assert(n->outcnt() == 0, "all uses must be deleted"); 1474 } 1475 1476 //------------------------------add_users_to_worklist-------------------------- 1477 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) { 1478 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1479 worklist.push(n->fast_out(i)); // Push on worklist 1480 } 1481 } 1482 1483 // Return counted loop Phi if as a counted loop exit condition, cmp 1484 // compares the induction variable with n 1485 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) { 1486 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1487 Node* bol = cmp->fast_out(i); 1488 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1489 Node* iff = bol->fast_out(i2); 1490 if (iff->is_BaseCountedLoopEnd()) { 1491 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd(); 1492 if (cle->limit() == n) { 1493 PhiNode* phi = cle->phi(); 1494 if (phi != nullptr) { 1495 return phi; 1496 } 1497 } 1498 } 1499 } 1500 } 1501 return nullptr; 1502 } 1503 1504 void PhaseIterGVN::add_users_to_worklist(Node *n) { 1505 add_users_to_worklist0(n, _worklist); 1506 1507 Unique_Node_List& worklist = _worklist; 1508 // Move users of node to worklist 1509 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1510 Node* use = n->fast_out(i); // Get use 1511 add_users_of_use_to_worklist(n, use, worklist); 1512 } 1513 } 1514 1515 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) { 1516 if(use->is_Multi() || // Multi-definer? Push projs on worklist 1517 use->is_Store() ) // Enable store/load same address 1518 add_users_to_worklist0(use, worklist); 1519 1520 // If we changed the receiver type to a call, we need to revisit 1521 // the Catch following the call. It's looking for a non-null 1522 // receiver to know when to enable the regular fall-through path 1523 // in addition to the NullPtrException path. 1524 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1525 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control); 1526 if (p != nullptr) { 1527 add_users_to_worklist0(p, worklist); 1528 } 1529 } 1530 1531 // AndLNode::Ideal folds GraphKit::mark_word_test patterns. Give it a chance to run. 1532 if (n->is_Load() && use->is_Phi()) { 1533 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1534 Node* u = use->fast_out(i); 1535 if (u->Opcode() == Op_AndL) { 1536 worklist.push(u); 1537 } 1538 } 1539 } 1540 1541 uint use_op = use->Opcode(); 1542 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1543 add_users_to_worklist0(use, worklist); // Put Bool on worklist 1544 if (use->outcnt() > 0) { 1545 Node* bol = use->raw_out(0); 1546 if (bol->outcnt() > 0) { 1547 Node* iff = bol->raw_out(0); 1548 if (iff->outcnt() == 2) { 1549 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1550 // phi merging either 0 or 1 onto the worklist 1551 Node* ifproj0 = iff->raw_out(0); 1552 Node* ifproj1 = iff->raw_out(1); 1553 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1554 Node* region0 = ifproj0->raw_out(0); 1555 Node* region1 = ifproj1->raw_out(0); 1556 if( region0 == region1 ) 1557 add_users_to_worklist0(region0, worklist); 1558 } 1559 } 1560 } 1561 } 1562 if (use_op == Op_CmpI || use_op == Op_CmpL) { 1563 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n); 1564 if (phi != nullptr) { 1565 // Input to the cmp of a loop exit check has changed, thus 1566 // the loop limit may have changed, which can then change the 1567 // range values of the trip-count Phi. 1568 worklist.push(phi); 1569 } 1570 } 1571 if (use_op == Op_CmpI) { 1572 Node* cmp = use; 1573 Node* in1 = cmp->in(1); 1574 Node* in2 = cmp->in(2); 1575 // Notify CmpI / If pattern from CastIINode::Value (left pattern). 1576 // Must also notify if in1 is modified and possibly turns into X (right pattern). 1577 // 1578 // in1 in2 in1 in2 1579 // | | | | 1580 // +--- | --+ | | 1581 // | | | | | 1582 // CmpINode | CmpINode 1583 // | | | 1584 // BoolNode | BoolNode 1585 // | | OR | 1586 // IfNode | IfNode 1587 // | | | 1588 // IfProj | IfProj X 1589 // | | | | 1590 // CastIINode CastIINode 1591 // 1592 if (in1 != in2) { // if they are equal, the CmpI can fold them away 1593 if (in1 == n) { 1594 // in1 modified -> could turn into X -> do traversal based on right pattern. 1595 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) { 1596 Node* bol = cmp->fast_out(i2); // For each Bool 1597 if (bol->is_Bool()) { 1598 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) { 1599 Node* iff = bol->fast_out(i3); // For each If 1600 if (iff->is_If()) { 1601 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) { 1602 Node* if_proj = iff->fast_out(i4); // For each IfProj 1603 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs"); 1604 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) { 1605 Node* castii = if_proj->fast_out(i5); // For each CastII 1606 if (castii->is_CastII() && 1607 castii->as_CastII()->carry_dependency()) { 1608 worklist.push(castii); 1609 } 1610 } 1611 } 1612 } 1613 } 1614 } 1615 } 1616 } else { 1617 // Only in2 modified -> can assume X == in2 (left pattern). 1618 assert(n == in2, "only in2 modified"); 1619 // Find all CastII with input in1. 1620 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) { 1621 Node* castii = in1->fast_out(j); 1622 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) { 1623 // Find If. 1624 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) { 1625 Node* ifnode = castii->in(0)->in(0); 1626 // Check that if connects to the cmp 1627 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) { 1628 worklist.push(castii); 1629 } 1630 } 1631 } 1632 } 1633 } 1634 } 1635 } 1636 } 1637 1638 // Inline type nodes can have other inline types as users. If an input gets 1639 // updated, make sure that inline type users get a chance for optimization. 1640 if (use->is_InlineType()) { 1641 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1642 Node* u = use->fast_out(i2); 1643 if (u->is_InlineType()) 1644 worklist.push(u); 1645 } 1646 } 1647 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast" 1648 // Patterns: 1649 // ConstraintCast+ -> Sub 1650 // ConstraintCast+ -> Phi 1651 // ConstraintCast+ -> Xor 1652 if (use->is_ConstraintCast()) { 1653 auto push_the_uses_to_worklist = [&](Node* n){ 1654 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) { 1655 worklist.push(n); 1656 } 1657 }; 1658 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); }; 1659 use->visit_uses(push_the_uses_to_worklist, is_boundary); 1660 } 1661 // If changed LShift inputs, check RShift users for useless sign-ext 1662 if (use_op == Op_LShiftI || use_op == Op_LShiftL) { 1663 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1664 Node* u = use->fast_out(i2); 1665 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL) 1666 worklist.push(u); 1667 } 1668 } 1669 // If changed LShift inputs, check And users for shift and mask (And) operation 1670 if (use_op == Op_LShiftI || use_op == Op_LShiftL) { 1671 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1672 Node* u = use->fast_out(i2); 1673 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) { 1674 worklist.push(u); 1675 } 1676 } 1677 } 1678 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1679 if (use_op == Op_AddI || use_op == Op_SubI) { 1680 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1681 Node* u = use->fast_out(i2); 1682 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1683 worklist.push(u); 1684 } 1685 } 1686 } 1687 // If changed AddP inputs: 1688 // - check Stores for loop invariant, and 1689 // - if the changed input is the offset, check constant-offset AddP users for 1690 // address expression flattening. 1691 if (use_op == Op_AddP) { 1692 bool offset_changed = n == use->in(AddPNode::Offset); 1693 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1694 Node* u = use->fast_out(i2); 1695 if (u->is_Mem()) { 1696 worklist.push(u); 1697 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) { 1698 worklist.push(u); 1699 } 1700 } 1701 } 1702 // If changed initialization activity, check dependent Stores 1703 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1704 InitializeNode* init = use->as_Allocate()->initialization(); 1705 if (init != nullptr) { 1706 Node* imem = init->proj_out_or_null(TypeFunc::Memory); 1707 if (imem != nullptr) add_users_to_worklist0(imem, worklist); 1708 } 1709 } 1710 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead. 1711 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn 1712 // to guarantee the change is not missed. 1713 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) { 1714 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control); 1715 if (p != nullptr) { 1716 add_users_to_worklist0(p, worklist); 1717 } 1718 } 1719 1720 if (use_op == Op_Initialize) { 1721 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory); 1722 if (imem != nullptr) add_users_to_worklist0(imem, worklist); 1723 } 1724 // Loading the java mirror from a Klass requires two loads and the type 1725 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1726 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1727 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1728 bool has_load_barrier_nodes = bs->has_load_barrier_nodes(); 1729 1730 if (use_op == Op_CastP2X) { 1731 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1732 Node* u = use->fast_out(i2); 1733 if (u->Opcode() == Op_AndX) { 1734 worklist.push(u); 1735 } 1736 } 1737 } 1738 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) { 1739 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1740 Node* u = use->fast_out(i2); 1741 const Type* ut = u->bottom_type(); 1742 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) { 1743 if (has_load_barrier_nodes) { 1744 // Search for load barriers behind the load 1745 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1746 Node* b = u->fast_out(i3); 1747 if (bs->is_gc_barrier_node(b)) { 1748 worklist.push(b); 1749 } 1750 } 1751 } 1752 worklist.push(u); 1753 } 1754 } 1755 } 1756 // Give CallStaticJavaNode::remove_useless_allocation a chance to run 1757 if (use->is_Region()) { 1758 Node* c = use; 1759 do { 1760 c = c->unique_ctrl_out_or_null(); 1761 } while (c != nullptr && c->is_Region()); 1762 if (c != nullptr && c->is_CallStaticJava() && c->as_CallStaticJava()->uncommon_trap_request() != 0) { 1763 worklist.push(c); 1764 } 1765 } 1766 if (use->Opcode() == Op_OpaqueZeroTripGuard) { 1767 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared"); 1768 if (use->outcnt() == 1) { 1769 Node* cmp = use->unique_out(); 1770 worklist.push(cmp); 1771 } 1772 } 1773 if (use->Opcode() == Op_AddX) { 1774 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1775 Node* u = use->fast_out(i2); 1776 if (u->Opcode() == Op_CastX2P) { 1777 worklist.push(u); 1778 } 1779 } 1780 } 1781 } 1782 1783 /** 1784 * Remove the speculative part of all types that we know of 1785 */ 1786 void PhaseIterGVN::remove_speculative_types() { 1787 assert(UseTypeSpeculation, "speculation is off"); 1788 for (uint i = 0; i < _types.Size(); i++) { 1789 const Type* t = _types.fast_lookup(i); 1790 if (t != nullptr) { 1791 _types.map(i, t->remove_speculative()); 1792 } 1793 } 1794 _table.check_no_speculative_types(); 1795 } 1796 1797 // Check if the type of a divisor of a Div or Mod node includes zero. 1798 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const { 1799 switch (n->Opcode()) { 1800 case Op_DivI: 1801 case Op_ModI: 1802 case Op_UDivI: 1803 case Op_UModI: { 1804 // Type of divisor includes 0? 1805 if (type(n->in(2)) == Type::TOP) { 1806 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations. 1807 return false; 1808 } 1809 const TypeInt* type_divisor = type(n->in(2))->is_int(); 1810 return (type_divisor->_hi < 0 || type_divisor->_lo > 0); 1811 } 1812 case Op_DivL: 1813 case Op_ModL: 1814 case Op_UDivL: 1815 case Op_UModL: { 1816 // Type of divisor includes 0? 1817 if (type(n->in(2)) == Type::TOP) { 1818 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations. 1819 return false; 1820 } 1821 const TypeLong* type_divisor = type(n->in(2))->is_long(); 1822 return (type_divisor->_hi < 0 || type_divisor->_lo > 0); 1823 } 1824 } 1825 return true; 1826 } 1827 1828 //============================================================================= 1829 #ifndef PRODUCT 1830 uint PhaseCCP::_total_invokes = 0; 1831 uint PhaseCCP::_total_constants = 0; 1832 #endif 1833 //------------------------------PhaseCCP--------------------------------------- 1834 // Conditional Constant Propagation, ala Wegman & Zadeck 1835 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1836 NOT_PRODUCT( clear_constants(); ) 1837 assert( _worklist.size() == 0, "" ); 1838 analyze(); 1839 } 1840 1841 #ifndef PRODUCT 1842 //------------------------------~PhaseCCP-------------------------------------- 1843 PhaseCCP::~PhaseCCP() { 1844 inc_invokes(); 1845 _total_constants += count_constants(); 1846 } 1847 #endif 1848 1849 1850 #ifdef ASSERT 1851 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) { 1852 if (tnew->meet(told) != tnew->remove_speculative()) { 1853 n->dump(3); 1854 tty->print("told = "); told->dump(); tty->cr(); 1855 tty->print("tnew = "); tnew->dump(); tty->cr(); 1856 fatal("Not monotonic"); 1857 } 1858 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases"); 1859 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases"); 1860 } 1861 #endif //ASSERT 1862 1863 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until 1864 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time 1865 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren 1866 // or nodes even further down back to the worklist because their type could change as a result of the current type 1867 // change). 1868 void PhaseCCP::analyze() { 1869 // Initialize all types to TOP, optimistic analysis 1870 for (uint i = 0; i < C->unique(); i++) { 1871 _types.map(i, Type::TOP); 1872 } 1873 1874 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()". 1875 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after 1876 // Compile is over. The local arena gets de-allocated at the end of its scope. 1877 ResourceArea local_arena(mtCompiler); 1878 Unique_Node_List worklist(&local_arena); 1879 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);) 1880 1881 // Push root onto worklist 1882 worklist.push(C->root()); 1883 1884 assert(_root_and_safepoints.size() == 0, "must be empty (unused)"); 1885 _root_and_safepoints.push(C->root()); 1886 1887 // Pull from worklist; compute new value; push changes out. 1888 // This loop is the meat of CCP. 1889 while (worklist.size() != 0) { 1890 Node* n = fetch_next_node(worklist); 1891 DEBUG_ONLY(worklist_verify.push(n);) 1892 if (n->is_SafePoint()) { 1893 // Make sure safepoints are processed by PhaseCCP::transform even if they are 1894 // not reachable from the bottom. Otherwise, infinite loops would be removed. 1895 _root_and_safepoints.push(n); 1896 } 1897 const Type* new_type = n->Value(this); 1898 if (new_type != type(n)) { 1899 DEBUG_ONLY(verify_type(n, new_type, type(n));) 1900 dump_type_and_node(n, new_type); 1901 set_type(n, new_type); 1902 push_child_nodes_to_worklist(worklist, n); 1903 } 1904 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) { 1905 // Keep track of Type nodes to kill CFG paths that use Type 1906 // nodes that become dead. 1907 _maybe_top_type_nodes.push(n); 1908 } 1909 } 1910 DEBUG_ONLY(verify_analyze(worklist_verify);) 1911 } 1912 1913 #ifdef ASSERT 1914 // For every node n on verify list, check if type(n) == n->Value() 1915 // We have a list of exceptions, see comments in verify_node_value. 1916 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) { 1917 bool failure = false; 1918 while (worklist_verify.size()) { 1919 Node* n = worklist_verify.pop(); 1920 failure |= verify_node_value(n); 1921 } 1922 // If we get this assert, check why the reported nodes were not processed again in CCP. 1923 // We should either make sure that these nodes are properly added back to the CCP worklist 1924 // in PhaseCCP::push_child_nodes_to_worklist() to update their type or add an exception 1925 // in the verification code above if that is not possible for some reason (like Load nodes). 1926 assert(!failure, "PhaseCCP not at fixpoint: analysis result may be unsound."); 1927 } 1928 #endif 1929 1930 // Fetch next node from worklist to be examined in this iteration. 1931 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) { 1932 if (StressCCP) { 1933 return worklist.remove(C->random() % worklist.size()); 1934 } else { 1935 return worklist.pop(); 1936 } 1937 } 1938 1939 #ifndef PRODUCT 1940 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) { 1941 if (TracePhaseCCP) { 1942 t->dump(); 1943 do { 1944 tty->print("\t"); 1945 } while (tty->position() < 16); 1946 n->dump(); 1947 } 1948 } 1949 #endif 1950 1951 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes 1952 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result 1953 // of the new type of 'n'. Push these nodes to the worklist. 1954 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const { 1955 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1956 Node* use = n->fast_out(i); 1957 push_if_not_bottom_type(worklist, use); 1958 push_more_uses(worklist, n, use); 1959 } 1960 } 1961 1962 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const { 1963 if (n->bottom_type() != type(n)) { 1964 worklist.push(n); 1965 } 1966 } 1967 1968 // For some nodes, we need to propagate the type change to grandchildren or even further down. 1969 // Add them back to the worklist. 1970 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const { 1971 push_phis(worklist, use); 1972 push_catch(worklist, use); 1973 push_cmpu(worklist, use); 1974 push_counted_loop_phi(worklist, parent, use); 1975 push_cast(worklist, use); 1976 push_loadp(worklist, use); 1977 push_and(worklist, parent, use); 1978 push_cast_ii(worklist, parent, use); 1979 push_opaque_zero_trip_guard(worklist, use); 1980 } 1981 1982 1983 // We must recheck Phis too if use is a Region. 1984 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const { 1985 if (use->is_Region()) { 1986 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1987 push_if_not_bottom_type(worklist, use->fast_out(i)); 1988 } 1989 } 1990 } 1991 1992 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a 1993 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path. 1994 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes. 1995 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) { 1996 if (use->is_Call()) { 1997 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1998 Node* proj = use->fast_out(i); 1999 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) { 2000 Node* catch_node = proj->find_out_with(Op_Catch); 2001 if (catch_node != nullptr) { 2002 worklist.push(catch_node); 2003 } 2004 } 2005 } 2006 } 2007 } 2008 2009 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately 2010 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed 2011 // and get wrong types otherwise. 2012 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const { 2013 uint use_op = use->Opcode(); 2014 if (use_op == Op_AddI || use_op == Op_SubI) { 2015 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 2016 Node* cmpu = use->fast_out(i); 2017 const uint cmpu_opcode = cmpu->Opcode(); 2018 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) { 2019 // Got a CmpU or CmpU3 which might need the new type information from node n. 2020 push_if_not_bottom_type(worklist, cmpu); 2021 } 2022 } 2023 } 2024 } 2025 2026 // If n is used in a counted loop exit condition, then the type of the counted loop's Phi depends on the type of 'n'. 2027 // Seem PhiNode::Value(). 2028 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) { 2029 uint use_op = use->Opcode(); 2030 if (use_op == Op_CmpI || use_op == Op_CmpL) { 2031 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent); 2032 if (phi != nullptr) { 2033 worklist.push(phi); 2034 } 2035 } 2036 } 2037 2038 void PhaseCCP::push_cast(Unique_Node_List& worklist, const Node* use) { 2039 uint use_op = use->Opcode(); 2040 if (use_op == Op_CastP2X) { 2041 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 2042 Node* u = use->fast_out(i2); 2043 if (u->Opcode() == Op_AndX) { 2044 worklist.push(u); 2045 } 2046 } 2047 } 2048 } 2049 2050 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'. 2051 // See LoadNode::Value(). 2052 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const { 2053 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2(); 2054 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes(); 2055 2056 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) { 2057 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 2058 Node* loadp = use->fast_out(i); 2059 const Type* ut = loadp->bottom_type(); 2060 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) { 2061 if (has_load_barrier_nodes) { 2062 // Search for load barriers behind the load 2063 push_load_barrier(worklist, barrier_set, loadp); 2064 } 2065 worklist.push(loadp); 2066 } 2067 } 2068 } 2069 } 2070 2071 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) { 2072 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 2073 Node* barrier_node = use->fast_out(i); 2074 if (barrier_set->is_gc_barrier_node(barrier_node)) { 2075 worklist.push(barrier_node); 2076 } 2077 } 2078 } 2079 2080 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint. 2081 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift 2082 // value changed. 2083 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const { 2084 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type()); 2085 uint use_op = use->Opcode(); 2086 if ( 2087 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And 2088 (parent_type != nullptr && parent_type->is_con()) || 2089 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And 2090 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) { 2091 2092 auto push_and_uses_to_worklist = [&](Node* n) { 2093 uint opc = n->Opcode(); 2094 if (opc == Op_AndI || opc == Op_AndL) { 2095 push_if_not_bottom_type(worklist, n); 2096 } 2097 }; 2098 auto is_boundary = [](Node* n) { 2099 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L); 2100 }; 2101 use->visit_uses(push_and_uses_to_worklist, is_boundary); 2102 } 2103 } 2104 2105 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is 2106 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the 2107 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it 2108 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input 2109 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case. 2110 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const { 2111 if (use->Opcode() == Op_CmpI && use->in(2) == parent) { 2112 Node* other_cmp_input = use->in(1); 2113 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) { 2114 Node* cast_ii = other_cmp_input->fast_out(i); 2115 if (cast_ii->is_CastII()) { 2116 push_if_not_bottom_type(worklist, cast_ii); 2117 } 2118 } 2119 } 2120 } 2121 2122 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const { 2123 if (use->Opcode() == Op_OpaqueZeroTripGuard) { 2124 push_if_not_bottom_type(worklist, use->unique_out()); 2125 } 2126 } 2127 2128 //------------------------------do_transform----------------------------------- 2129 // Top level driver for the recursive transformer 2130 void PhaseCCP::do_transform() { 2131 // Correct leaves of new-space Nodes; they point to old-space. 2132 C->set_root( transform(C->root())->as_Root() ); 2133 assert( C->top(), "missing TOP node" ); 2134 assert( C->root(), "missing root" ); 2135 } 2136 2137 //------------------------------transform-------------------------------------- 2138 // Given a Node in old-space, clone him into new-space. 2139 // Convert any of his old-space children into new-space children. 2140 Node *PhaseCCP::transform( Node *n ) { 2141 assert(n->is_Root(), "traversal must start at root"); 2142 assert(_root_and_safepoints.member(n), "root (n) must be in list"); 2143 2144 ResourceMark rm; 2145 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless). 2146 Node_List node_map; 2147 // Pre-allocate to avoid frequent realloc 2148 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1); 2149 // track all visited nodes, so that we can remove the complement 2150 Unique_Node_List useful; 2151 2152 if (KillPathsReachableByDeadTypeNode) { 2153 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) { 2154 Node* type_node = _maybe_top_type_nodes.at(i); 2155 if (type(type_node) == Type::TOP) { 2156 ResourceMark rm; 2157 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp"); 2158 } 2159 } 2160 } else { 2161 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes"); 2162 } 2163 2164 // Initialize the traversal. 2165 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case, 2166 // the logic below doesn't follow any path from Root to the loop body: there's at least one such path but it's proven 2167 // never taken (its type is TOP). As a consequence the node on the exit path that's input to Root (let's call it n) is 2168 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works 2169 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that 2170 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root 2171 // and all safepoints. 2172 for (uint i = 0; i < _root_and_safepoints.size(); ++i) { 2173 Node* nn = _root_and_safepoints.at(i); 2174 Node* new_node = node_map[nn->_idx]; 2175 assert(new_node == nullptr, ""); 2176 new_node = transform_once(nn); // Check for constant 2177 node_map.map(nn->_idx, new_node); // Flag as having been cloned 2178 transform_stack.push(new_node); // Process children of cloned node 2179 useful.push(new_node); 2180 } 2181 2182 while (transform_stack.is_nonempty()) { 2183 Node* clone = transform_stack.pop(); 2184 uint cnt = clone->req(); 2185 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 2186 Node *input = clone->in(i); 2187 if( input != nullptr ) { // Ignore nulls 2188 Node *new_input = node_map[input->_idx]; // Check for cloned input node 2189 if( new_input == nullptr ) { 2190 new_input = transform_once(input); // Check for constant 2191 node_map.map( input->_idx, new_input );// Flag as having been cloned 2192 transform_stack.push(new_input); // Process children of cloned node 2193 useful.push(new_input); 2194 } 2195 assert( new_input == clone->in(i), "insanity check"); 2196 } 2197 } 2198 } 2199 2200 // The above transformation might lead to subgraphs becoming unreachable from the 2201 // bottom while still being reachable from the top. As a result, nodes in that 2202 // subgraph are not transformed and their bottom types are not updated, leading to 2203 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in 2204 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common 2205 // because their address type is inconsistent. Therefore, we aggressively remove 2206 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance 2207 // to remove them anyway. 2208 if (C->cached_top_node()) { 2209 useful.push(C->cached_top_node()); 2210 } 2211 C->update_dead_node_list(useful); 2212 remove_useless_nodes(useful.member_set()); 2213 _worklist.remove_useless_nodes(useful.member_set()); 2214 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints); 2215 2216 Node* new_root = node_map[n->_idx]; 2217 assert(new_root->is_Root(), "transformed root node must be a root node"); 2218 return new_root; 2219 } 2220 2221 //------------------------------transform_once--------------------------------- 2222 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 2223 Node *PhaseCCP::transform_once( Node *n ) { 2224 const Type *t = type(n); 2225 // Constant? Use constant Node instead 2226 if( t->singleton() ) { 2227 Node *nn = n; // Default is to return the original constant 2228 if( t == Type::TOP ) { 2229 // cache my top node on the Compile instance 2230 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) { 2231 C->set_cached_top_node(ConNode::make(Type::TOP)); 2232 set_type(C->top(), Type::TOP); 2233 } 2234 nn = C->top(); 2235 } 2236 if( !n->is_Con() ) { 2237 if( t != Type::TOP ) { 2238 nn = makecon(t); // ConNode::make(t); 2239 NOT_PRODUCT( inc_constants(); ) 2240 } else if( n->is_Region() ) { // Unreachable region 2241 // Note: nn == C->top() 2242 n->set_req(0, nullptr); // Cut selfreference 2243 bool progress = true; 2244 uint max = n->outcnt(); 2245 DUIterator i; 2246 while (progress) { 2247 progress = false; 2248 // Eagerly remove dead phis to avoid phis copies creation. 2249 for (i = n->outs(); n->has_out(i); i++) { 2250 Node* m = n->out(i); 2251 if (m->is_Phi()) { 2252 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 2253 replace_node(m, nn); 2254 if (max != n->outcnt()) { 2255 progress = true; 2256 i = n->refresh_out_pos(i); 2257 max = n->outcnt(); 2258 } 2259 } 2260 } 2261 } 2262 } 2263 replace_node(n,nn); // Update DefUse edges for new constant 2264 } 2265 return nn; 2266 } 2267 2268 // If x is a TypeNode, capture any more-precise type permanently into Node 2269 if (t != n->bottom_type()) { 2270 hash_delete(n); // changing bottom type may force a rehash 2271 n->raise_bottom_type(t); 2272 _worklist.push(n); // n re-enters the hash table via the worklist 2273 } 2274 2275 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks 2276 switch( n->Opcode() ) { 2277 case Op_CallStaticJava: // Give post-parse call devirtualization a chance 2278 case Op_CallDynamicJava: 2279 case Op_FastLock: // Revisit FastLocks for lock coarsening 2280 case Op_If: 2281 case Op_CountedLoopEnd: 2282 case Op_Region: 2283 case Op_Loop: 2284 case Op_CountedLoop: 2285 case Op_Conv2B: 2286 case Op_Opaque1: 2287 _worklist.push(n); 2288 break; 2289 default: 2290 break; 2291 } 2292 2293 return n; 2294 } 2295 2296 //---------------------------------saturate------------------------------------ 2297 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 2298 const Type* limit_type) const { 2299 const Type* wide_type = new_type->widen(old_type, limit_type); 2300 if (wide_type != new_type) { // did we widen? 2301 // If so, we may have widened beyond the limit type. Clip it back down. 2302 new_type = wide_type->filter(limit_type); 2303 } 2304 return new_type; 2305 } 2306 2307 //------------------------------print_statistics------------------------------- 2308 #ifndef PRODUCT 2309 void PhaseCCP::print_statistics() { 2310 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 2311 } 2312 #endif 2313 2314 2315 //============================================================================= 2316 #ifndef PRODUCT 2317 uint PhasePeephole::_total_peepholes = 0; 2318 #endif 2319 //------------------------------PhasePeephole---------------------------------- 2320 // Conditional Constant Propagation, ala Wegman & Zadeck 2321 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 2322 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 2323 NOT_PRODUCT( clear_peepholes(); ) 2324 } 2325 2326 #ifndef PRODUCT 2327 //------------------------------~PhasePeephole--------------------------------- 2328 PhasePeephole::~PhasePeephole() { 2329 _total_peepholes += count_peepholes(); 2330 } 2331 #endif 2332 2333 //------------------------------transform-------------------------------------- 2334 Node *PhasePeephole::transform( Node *n ) { 2335 ShouldNotCallThis(); 2336 return nullptr; 2337 } 2338 2339 //------------------------------do_transform----------------------------------- 2340 void PhasePeephole::do_transform() { 2341 bool method_name_not_printed = true; 2342 2343 // Examine each basic block 2344 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 2345 Block* block = _cfg.get_block(block_number); 2346 bool block_not_printed = true; 2347 2348 for (bool progress = true; progress;) { 2349 progress = false; 2350 // block->end_idx() not valid after PhaseRegAlloc 2351 uint end_index = block->number_of_nodes(); 2352 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) { 2353 Node *n = block->get_node(instruction_index); 2354 if( n->is_Mach() ) { 2355 MachNode *m = n->as_Mach(); 2356 // check for peephole opportunities 2357 int result = m->peephole(block, instruction_index, &_cfg, _regalloc); 2358 if( result != -1 ) { 2359 #ifndef PRODUCT 2360 if( PrintOptoPeephole ) { 2361 // Print method, first time only 2362 if( C->method() && method_name_not_printed ) { 2363 C->method()->print_short_name(); tty->cr(); 2364 method_name_not_printed = false; 2365 } 2366 // Print this block 2367 if( Verbose && block_not_printed) { 2368 tty->print_cr("in block"); 2369 block->dump(); 2370 block_not_printed = false; 2371 } 2372 // Print the peephole number 2373 tty->print_cr("peephole number: %d", result); 2374 } 2375 inc_peepholes(); 2376 #endif 2377 // Set progress, start again 2378 progress = true; 2379 break; 2380 } 2381 } 2382 } 2383 } 2384 } 2385 } 2386 2387 //------------------------------print_statistics------------------------------- 2388 #ifndef PRODUCT 2389 void PhasePeephole::print_statistics() { 2390 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2391 } 2392 #endif 2393 2394 2395 //============================================================================= 2396 //------------------------------set_req_X-------------------------------------- 2397 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2398 assert( is_not_dead(n), "can not use dead node"); 2399 #ifdef ASSERT 2400 if (igvn->hash_find(this) == this) { 2401 tty->print_cr("Need to remove from hash before changing edges"); 2402 this->dump(1); 2403 tty->print_cr("Set at i = %d", i); 2404 n->dump(); 2405 assert(false, "Need to remove from hash before changing edges"); 2406 } 2407 #endif 2408 Node *old = in(i); 2409 set_req(i, n); 2410 2411 // old goes dead? 2412 if( old ) { 2413 switch (old->outcnt()) { 2414 case 0: 2415 // Put into the worklist to kill later. We do not kill it now because the 2416 // recursive kill will delete the current node (this) if dead-loop exists 2417 if (!old->is_top()) 2418 igvn->_worklist.push( old ); 2419 break; 2420 case 1: 2421 if( old->is_Store() || old->has_special_unique_user() ) 2422 igvn->add_users_to_worklist( old ); 2423 break; 2424 case 2: 2425 if( old->is_Store() ) 2426 igvn->add_users_to_worklist( old ); 2427 if( old->Opcode() == Op_Region ) 2428 igvn->_worklist.push(old); 2429 break; 2430 case 3: 2431 if( old->Opcode() == Op_Region ) { 2432 igvn->_worklist.push(old); 2433 igvn->add_users_to_worklist( old ); 2434 } 2435 break; 2436 default: 2437 break; 2438 } 2439 2440 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old); 2441 } 2442 } 2443 2444 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) { 2445 PhaseIterGVN* igvn = gvn->is_IterGVN(); 2446 if (igvn == nullptr) { 2447 set_req(i, n); 2448 return; 2449 } 2450 set_req_X(i, n, igvn); 2451 } 2452 2453 //-------------------------------replace_by----------------------------------- 2454 // Using def-use info, replace one node for another. Follow the def-use info 2455 // to all users of the OLD node. Then make all uses point to the NEW node. 2456 void Node::replace_by(Node *new_node) { 2457 assert(!is_top(), "top node has no DU info"); 2458 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2459 Node* use = last_out(i); 2460 uint uses_found = 0; 2461 for (uint j = 0; j < use->len(); j++) { 2462 if (use->in(j) == this) { 2463 if (j < use->req()) 2464 use->set_req(j, new_node); 2465 else use->set_prec(j, new_node); 2466 uses_found++; 2467 } 2468 } 2469 i -= uses_found; // we deleted 1 or more copies of this edge 2470 } 2471 } 2472 2473 //============================================================================= 2474 //----------------------------------------------------------------------------- 2475 void Type_Array::grow( uint i ) { 2476 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena"); 2477 if( !_max ) { 2478 _max = 1; 2479 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2480 _types[0] = nullptr; 2481 } 2482 uint old = _max; 2483 _max = next_power_of_2(i); 2484 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2485 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2486 } 2487 2488 //------------------------------dump------------------------------------------- 2489 #ifndef PRODUCT 2490 void Type_Array::dump() const { 2491 uint max = Size(); 2492 for( uint i = 0; i < max; i++ ) { 2493 if( _types[i] != nullptr ) { 2494 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2495 } 2496 } 2497 } 2498 #endif